Explainable AI for CVD Risk Prediction: An Ensemble Learning Approach with Clinical Interpretability

Abstract

This research created and tested an enhanced machine learning system for predicting cardiovascular risk, tackling three key data challenges: handling missing values, selecting relevant features, and managing imbalanced classes. The approach combined automatic data cleaning (using median/mode replacement), SMOTE for balancing classes, XG Boost-based feature selection, and polynomial features to model complex patterns. Four algorithms were assessed - optimized XG Boost, Random Forest, simplified Decision Tree, and a combined voting model - using 918 patient records. The ensemble method showed the best results (87.5% correct predictions, 0.92 AUC score), effectively identifying at-risk patients. Notable advancements were: (1) merging medical expertise with algorithmic feature selection, (2) boosting minority class detection by 22% using SMOTE with ensemble methods, and (3) creating standardized risk assessment aligned with medical protocols. The pipeline bridges critical gaps in CVD prediction by combining high accuracy with clinician-friendly explainability, offering a deployable tool for risk stratification.

Ähnliche Arbeiten